Physics-Assisted Deep Learning Denoising for Stabilized IMPULSED dMRI Microenvironment Parameter Fitting

Diffusion-weighted MRI (dMRI) is a powerful tool for quantifying cellular microenvironment parameters. However, the inherently low signal-to-noise ratio (SNR) of dMRI can compromise the accuracy and reliability of parameter estimation. This study proposes a physics-assisted deep learning (DL)-based denoising framework designed to enhance dMRI signal quality and improve the robustness of subsequent biophysical model fitting, with potential relevance to low-SNR settings such as clinical 1.5 T MRI acquisitions.

A dataset of paired noise-free and Rician-noise-corrupted dMRI signals was generated using the IMPULSED-dMRI signal model. Three denoising architectures were evaluated: Convolutional Neural Networks (CNN), Multilayer Perceptron (MLP), and Long Short-Term Memory (LSTM) networks. Denoised signals were then fitted to estimate cell diameter $d$ , intracellular volume fraction $V_{\text{in}}$ , and extracellular apparent diffusion coefficient $D_{ex}$ . Performance was assessed using signal-domain Mean Absolute Error (MAE), fitted-parameter MAE, and fitting failure rate in synthetic IMPULSED-dMRI data with known ground-truth parameters and in in vitro evaluation using HeLa and MC38 cell lines.

DL-based processing substantially improved dMRI signal denoising. The MLP and LSTM achieved similar performance, with the LSTM slightly better overall, and both outperformed the CNN. Averaged across all signals, the LSTM reduced denoising MAE from 5.97% to 1.58%. In the subsequent model fitting step, the LSTM produced modest reductions in parameter MAE, from 5.34 μm to 4.02 μm for $d$ , from 15.00% to 11.96% for $V_{in}$ , and from 0.77 to 0.63 μm² /ms for $D_{\text{ex}}$ . The dominant benefit was fitting stabilization, with the overall fitting failure rate reduced from 57.6% to 17.7%. In in vitro experiments, relative to experimental references, the LSTM reduced MAE from 2.1 μm to 0.4 μm for $d$ and from 7.8 to 5.4 percentage points for $V_{in}$ , while reducing the mean overall fitting failure rate from 17.7% to 0%.

The proposed framework improves dMRI signal quality and stabilizes subsequent IMPULSED-based microenvironmental parameter fitting. The primary value of this approach is improved fitting reliability under noisy dMRI conditions, with secondary gains in parameter accuracy, and it warrants further validation in heterogeneous tissues and in vivo datasets.